Re: [dev] [sbase] New branch for sbase+ubase
On 2024-03-14 11:49, Roberto E. Vargas Caballero wrote: The idea is that everything is portable. Let's take the example of ps. We can extract the non portable bits of ps into the libsys library (or whatever name we like), so the code in posix/ is portable. Ok, that is evidently better. Having in libsys/ all the nonportable bits is feasible and better than re-implementing each tool for each new platform. I think some tools such as sysctl are going to be way more complex than that though. Linux kindly provides /proc/sys, and the current sysctl merely replaces '.' by '/' and writes to the corresponding file. Such an implementation is impossible on OpenBSD. It does not provide any systematic way of mapping a "key" (such as "vm.drop_caches", or "vm.loadavg") to the underlying bit-array that uniquely identifies it for sysctl(2). We would need to write thousands of lines of OS-dependant code just to _parse_ the command-line. But sysctl _is_ fundamental for a working userland. Should we thus put it in linux/, or find a way to make it portable nonetheless? I don't get why we should have categories like "shell-only", or why we should have so high level of customization. Well, this is just a suggestion. I can see why it would be a bad idea to have such high level of customization. But I sill believe that those categories should be defined in the Makefile rather than in directories. It is _very_ easy to pick-and-choose which utilities to build with the master branch. It just requires to modify the BIN list, ie. remove tools that don't fit. With a directory-based separation, such a process would be more complex, and would require to modify each Makefile individually; which, arguably, is cumbersome. This is basically just overlapping lists of utilities, defined by standards, type of utility, etc. What is the reasoning for categories like "shell-only"? The categories you've outlined are very clear and make sense; but they aren't the only way of grouping tools. "shell-only" was just an example, but some tools only make sense when used in a terminal, for displaying data to the user. Sometimes, this is not needed (such as routers for example, where there is very little need for tools like cal, cols or clear). Maybe this it way too specific (and way too arbitrary), but it was just an example. But if we ever implement tools such as gzip or xz, ping or nc, etc. it _could_ be useful to have sets such as "compression", "net", that may overlap with other categories. This proposed scheme is quite inspired by the way busybox separates its utilities, but maybe it is wrong for this precise reason. I find the directory separation both in toybox and busybox quite useful and elegant, so that's why I'm suggesting a system that would allow for such separation in a simple way (Makefile lists). I put only linux because at this moment we only have linux specific tools, if we add OpenBSD tools then we can have the openbsd directory. Of course that it would mean that if you add linux and openbsd at the same time your build will fail, just don't do stupid things. - Allow for categories _inside the Makefile_. We could have something like: POSIXTOOLS = portable/ls unportable/$(OS)/ps.c ... MISCTOOLS = portable/sponge.c portable/cols.c portable/rev.c ... INTERACTIVETOOLS = portable/cols.c portable/clear.c ... ... BIN = $(POSIXTOOLS) $(MISCTOOLS) $(LINUXTOOLS) At this stage and with the configuration that we want to have I think we have to get rid of the monolithic approach and just define the directories that we want to include in our build: all: posix misc curses and every directory with simple Makefiles. As commented before, I think this is problematic for users who want more control on which tools they want to build. Also, I'm really not sure how you could get rid of the monolithic approach while still being able to build sbase-box. The current script-based method is quite good I think, except that it builds all the C files in the current directory. The Makefile recipe would need to provide the script with which files to build, ie. with lists of C files. The root Makefile has to have knowledge of the C files in each subdirectory, and pass it to mkbox (along with which libsys version to use). We would have a non-monolithic build for individual tools, but a monolithic one for sbase-box, which means duplicating lists of utilities. If we get rid of the monolithic approach, then what I propose (overlapping categories as lists of utilities in a Makefile) would be inapplicable. As commented, the idea is to have all the tools written with portable code. Tools in POSIX are implemented in all the systems, so they have a way to be implemented. At this moment I just began to classify the tools, that is not easy, and I didn't care about fixing the portability problems, but after deciding the categories then it should be the next step. Of course! I think some
Re: [dev] [sbase] New branch for sbase+ubase
Hi, On Wed, Mar 13, 2024 at 11:51:46PM +0100, Elie Le Vaillant wrote: > I think this layout has a few problems: > > - we lose the meaningful separation that the sbase/ubase layout allowed, > i.e the distinction > between what is portable and what isn't. This _could_ be included as > part of the Makefile > (PORTABLE set vs NONPORTABLE), but I think it is better to explicitly > separate in different > directories portable from non-portable tools. If we don't, we are making > sbase a Linux-only > project (the only implementation for ps becomes a Linux-only one, and we > would need to fork > sbase to make it available to other platforms; same for all the other > tools from ubase), which > I find a bit sad considering the goals of portability of sbase The idea is that everything is portable. Let's take the example of ps. We can extract the non portable bits of ps into the libsys library (or whatever name we like), so the code in posix/ is portable. In some cases it would mean to remove some functionalities or worst performance, but it is not a problem because we can keep in ubase the optimized versions. For example, sbase has a portable dd and ubase has a customized linux dd version. > - I think Mattias Andrée scheme is better than this one. With a > directory-base separation > (between categories of utilities, based on somewhat arbitrary factors > (standards, rather than > say, minimalness, or use-case, or platform or implementation, etc.)), we > cannot have overlapping > categories, which is quite problematic. For example, if we wished to > have a category for > "shell-only usage" (such as clear, or cols), we couldn't implement it in > an easy way, because > this category overlaps with misc/, curses-dummy/, and maybe others too. I don't get why we should have categories like "shell-only", or why we should have so high level of customization. Having some categories makes sense, but going to the level of user custom lists is too much in my opinion. The separation between posix/ and misc/ is there because may people raised the concern that they didn't use tools like sponge, tac or shuf ever, so it is questionable to include them, but as there are other people that use them we add the option to include them. The curses tools have the problem that include a strong dependency with a complex library, and it is desirable to skip them, because it can be very problematic to build them in some cases. Having the unix/ category is there for all the tools that were part of POSIX previously and/or were part of UNIX historically but for different reasons are not part of POSIX today. Having a POSIX only set of tools has the advantage of being a tool to check if your shell script is portable or not. What is the reasoning for categories like "shell-only"? > What I suggest to fix both problems: > - separate on the grounds of portability/nonportability. In other words, > something along the lines > of: >libutil/ >portable/ > ls.c > cols.c As commented before, all the tools must be portable, except of course linux specific tools. After talking with quinq he liked the idea because then tools for Openbsd can be easily added just adding a new openbsd directory. > This more or less reproduces the sbase/ubase separation, but allowing > future OSes to come in > the future rather than just Linux (so, for example, *maybe* OpenBSD I put only linux because at this moment we only have linux specific tools, if we add OpenBSD tools then we can have the openbsd directory. Of course that it would mean that if you add linux and openbsd at the same time your build will fail, just don't do stupid things. > - Allow for categories _inside the Makefile_. We could have something > like: >POSIXTOOLS = portable/ls unportable/$(OS)/ps.c ... >MISCTOOLS = portable/sponge.c portable/cols.c portable/rev.c ... >INTERACTIVETOOLS = portable/cols.c portable/clear.c ... >... >BIN = $(POSIXTOOLS) $(MISCTOOLS) $(LINUXTOOLS) At this stage and with the configuration that we want to have I think we have to get rid of the monolithic approach and just define the directories that we want to include in our build: all: posix misc curses and every directory with simple Makefiles. > This allows for such grouping, while also allowing overlapping > categories. This also doesn't > hinder the useful semantic sbase/ubase separation (which is especially > handy when working on > non-Linux OSes). I think overall that sbase/ubase is the most useful > distinction, so it should > be treated specially, and not as part of the build-system (but more as > part of the directory > organization). As commented, the idea is to have all the tools written with portable code. Tools in POSIX are implemented in all the systems, so they have a way to be implemented. At this moment I just began to classify the tools, that is
Re: [dev] [sbase] New branch for sbase+ubase
On 2024-03-13 11:27, Roberto E. Vargas Caballero wrote: I am thinking about the new layout, and for now I was considering something like: - posix (tools defined by POSIX) - misc (helpful tools not defined by POSIX). - linux (tools that only make sense in linux) - curses-terminfo (tools with dependency of terminfo, like clean and reset). - curses-dummy (curses tools implementation without using terminfo). - libutil (library with utility functions) - libsys (library with system dependant functions). I think this layout has a few problems: - we lose the meaningful separation that the sbase/ubase layout allowed, i.e the distinction between what is portable and what isn't. This _could_ be included as part of the Makefile (PORTABLE set vs NONPORTABLE), but I think it is better to explicitly separate in different directories portable from non-portable tools. If we don't, we are making sbase a Linux-only project (the only implementation for ps becomes a Linux-only one, and we would need to fork sbase to make it available to other platforms; same for all the other tools from ubase), which I find a bit sad considering the goals of portability of sbase - I think Mattias Andrée scheme is better than this one. With a directory-base separation (between categories of utilities, based on somewhat arbitrary factors (standards, rather than say, minimalness, or use-case, or platform or implementation, etc.)), we cannot have overlapping categories, which is quite problematic. For example, if we wished to have a category for "shell-only usage" (such as clear, or cols), we couldn't implement it in an easy way, because this category overlaps with misc/, curses-dummy/, and maybe others too. What I suggest to fix both problems: - separate on the grounds of portability/nonportability. In other words, something along the lines of: libutil/ portable/ ls.c cols.c unportable/ linux/ lsmod.c ps.c libsys/ maybe-other-os-in-the-future/ This more or less reproduces the sbase/ubase separation, but allowing future OSes to come in the future rather than just Linux (so, for example, *maybe* OpenBSD could have some tools in the future. This is a suggestion). - Allow for categories _inside the Makefile_. We could have something like: POSIXTOOLS = portable/ls unportable/$(OS)/ps.c ... MISCTOOLS = portable/sponge.c portable/cols.c portable/rev.c ... INTERACTIVETOOLS = portable/cols.c portable/clear.c ... ... BIN = $(POSIXTOOLS) $(MISCTOOLS) $(LINUXTOOLS) This allows for such grouping, while also allowing overlapping categories. This also doesn't hinder the useful semantic sbase/ubase separation (which is especially handy when working on non-Linux OSes). I think overall that sbase/ubase is the most useful distinction, so it should be treated specially, and not as part of the build-system (but more as part of the directory organization). Possible drawbacks which I've thought about could be: - This requires to occasionally write and update different, overlapping lists of tools. I believe this is not too much of an issue. Adding new items should be quite easy, and you've already done a form a grouping which we could use for this proposed alternative layout. This task could also be aided by the way busybox groups its different utilities, and the different standards and packages listed on the toybox website. - The Makefile would be a bit more complex. If we truly add support for multiple platforms (which is a mere suggestion), it could become a bit complex (as in: platform-specific bits of Makefile to properly compile libsys, or platform-specific utilities (lsmod, sysctl...)). If we don't, and just keep a portable/linux separation at the directory-level, we still need a bit more complexity (defining tool categories, properly distinguishing between libutil and libsys at compile-time, and only for platform-specific tools, etc.) than what we currently have, but this complexity would be very similar to the one we would have with your proposed layout. Overall, I think that the benefits of the proposed alternative layout (sbase/ubase primary distinction, reflected at the directory level; and Makefile-specific lists as secondary distinction, only reflected at the build-system level, which allows very great flexibility) outweigh the inconveniences. Moreover, I think that your layout has issues, such as the fact that sbase becomes a de-facto Linux-only project, considering that tools whose implementation are Linux-specific (ps, mknod, etc.) are not treated differently from tools which are inherently portable (sponge, sed...), or the fact that the categories you've made so far forbid any kind of
[dev] [sbase] New branch for sbase+ubase
Hi, I have pushed a new branch called ubase-merge that has all the files from ubase in the directory ubase of the root directory in the repository. You can still see the full history of a file in that directory using something like git log --follow. I am thinking about the new layout, and for now I was considering something like: - posix (tools defined by POSIX) - misc (helpful tools not defined by POSIX). - linux (tools that only make sense in linux) - curses-terminfo (tools with dependency of terminfo, like clean and reset). - curses-dummy (curses tools implementation without using terminfo). - libutil (library with utility functions) - libsys (library with system dependant functions). The separation between curses and curses-dummy is because we have currently some curses tools implemented using hardcoded sequences instead of using terminfo to locate the correct sequences for the terminal used. While this would work in the majority of terminal emulators (because almost all of them emulate vt100 compatible emulatros) it would not work in many cases. It is a good idea to keep them because it reduces the dependencies of the project, and it makes easier to bootstrap systems. But I think we can also implement the correct solutions and select what is the option at build time. it was commented to have functions that isolated the system dependencies between different systems, and I thought that instead of adding them to libutil it was better to define a new library only for that purpose. Please, give your opinion. Regards,